CN116504005A - Perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM - Google Patents
Perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM Download PDFInfo
- Publication number
- CN116504005A CN116504005A CN202310512549.8A CN202310512549A CN116504005A CN 116504005 A CN116504005 A CN 116504005A CN 202310512549 A CN202310512549 A CN 202310512549A CN 116504005 A CN116504005 A CN 116504005A
- Authority
- CN
- China
- Prior art keywords
- layer
- data
- cdil
- lstm
- output
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 39
- 239000013307 optical fiber Substances 0.000 claims abstract description 32
- 230000003287 optical effect Effects 0.000 claims abstract description 22
- 238000001514 detection method Methods 0.000 claims abstract description 10
- 238000012706 support-vector machine Methods 0.000 claims abstract description 9
- 238000005516 engineering process Methods 0.000 claims abstract description 8
- 239000000835 fiber Substances 0.000 claims abstract description 8
- 230000009467 reduction Effects 0.000 claims abstract description 6
- 238000010606 normalization Methods 0.000 claims description 16
- 238000012549 training Methods 0.000 claims description 12
- 125000004122 cyclic group Chemical group 0.000 claims description 10
- 238000005070 sampling Methods 0.000 claims description 7
- 239000013598 vector Substances 0.000 claims description 7
- 230000009194 climbing Effects 0.000 claims description 6
- 238000000227 grinding Methods 0.000 claims description 5
- 230000003213 activating effect Effects 0.000 claims description 4
- 230000010339 dilation Effects 0.000 claims description 4
- 239000000284 extract Substances 0.000 claims description 4
- 238000011176 pooling Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 230000009466 transformation Effects 0.000 claims description 4
- 230000004931 aggregating effect Effects 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 238000012795 verification Methods 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract description 8
- 238000007781 pre-processing Methods 0.000 abstract description 5
- 238000010276 construction Methods 0.000 abstract description 2
- 230000011218 segmentation Effects 0.000 abstract description 2
- 239000010410 layer Substances 0.000 description 69
- 238000013527 convolutional neural network Methods 0.000 description 19
- 230000006870 function Effects 0.000 description 11
- 230000008569 process Effects 0.000 description 11
- 238000013528 artificial neural network Methods 0.000 description 10
- 238000004422 calculation algorithm Methods 0.000 description 10
- 238000013135 deep learning Methods 0.000 description 9
- 230000004927 fusion Effects 0.000 description 7
- 230000004913 activation Effects 0.000 description 6
- 230000002457 bidirectional effect Effects 0.000 description 6
- 238000010586 diagram Methods 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 4
- 238000010801 machine learning Methods 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 230000001427 coherent effect Effects 0.000 description 3
- 230000007547 defect Effects 0.000 description 3
- 238000000605 extraction Methods 0.000 description 3
- 238000002679 ablation Methods 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 230000007246 mechanism Effects 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 238000003062 neural network model Methods 0.000 description 2
- 238000007637 random forest analysis Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 238000007792 addition Methods 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000008033 biological extinction Effects 0.000 description 1
- 239000000872 buffer Substances 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000005520 cutting process Methods 0.000 description 1
- 238000003066 decision tree Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000004880 explosion Methods 0.000 description 1
- 238000007499 fusion processing Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 230000006386 memory function Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000001208 nuclear magnetic resonance pulse sequence Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 239000011241 protective layer Substances 0.000 description 1
- 238000011946 reduction process Methods 0.000 description 1
- 230000003014 reinforcing effect Effects 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000010079 rubber tapping Methods 0.000 description 1
- 230000035945 sensitivity Effects 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G08—SIGNALLING
- G08B—SIGNALLING OR CALLING SYSTEMS; ORDER TELEGRAPHS; ALARM SYSTEMS
- G08B13/00—Burglar, theft or intruder alarms
- G08B13/02—Mechanical actuation
- G08B13/12—Mechanical actuation by the breaking or disturbance of stretched cords or wires
- G08B13/122—Mechanical actuation by the breaking or disturbance of stretched cords or wires for a perimeter fence
- G08B13/124—Mechanical actuation by the breaking or disturbance of stretched cords or wires for a perimeter fence with the breaking or disturbance being optically detected, e.g. optical fibers in the perimeter fence
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Image Analysis (AREA)
- Burglar Alarm Systems (AREA)
Abstract
The invention relates to the technical field of perimeter security, in particular to a perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM, which comprises the following steps: s1, selecting an optical fiber distributed vibration or acoustic wave sensing system based on a phase-sensitive optical time domain reflection technology for system construction; s2, adopting a single-mode fiber as a detection optical cable to be laid along with the perimeter fence, and collecting data of an event by personnel; s3, marking the data after the data noise reduction treatment, wherein each event corresponds to a corresponding label, and constructing a data set; s4, inputting the single-channel time domain data into the CDIL layer, fusing the result output by the Bi-LSTM layer with the result output by the CDIL layer, and identifying and classifying the event to be detected through a support vector machine. The method is simple in preprocessing operation, only needs wavelet denoising and data segmentation of the data, does not need a large number of preprocessing operations, ensures the effectiveness and the authenticity of the data, and ensures the accuracy of the identification effect.
Description
Technical Field
The invention relates to the technical field of perimeter security, in particular to a perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM.
Background
The traditional perimeter security system monitors the intrusion behavior mainly through modes such as video monitoring, infrared correlation, electronic fence and the like, but the traditional security solutions have the defects of poor electromagnetic interference resistance, small detection range, high maintenance cost and the like. The optical fiber distributed vibration sensing technology has unique advantages in the fields with high requirements on safety and maintainability in long distance and large range by the characteristics of continuity, economy, safety and the like, and has been widely applied to perimeter security monitoring projects such as long-distance pipelines, border lines and the like.
Before deep learning develops, initially, an industry and an expert perform optical fiber distributed vibration sensing signal mode identification through a machine learning algorithm, a Support Vector Machine (SVM) algorithm is most widely applied, and certain classification effects are achieved through the machine learning algorithms such as an error Back Propagation (BP) algorithm, a decision tree, a Random Forest (RF), an Extreme Learning Machine (ELM), a random vector function chain (RVFL), a nearest neighbor classification network (KNN) and the like. In the Chinese patent specification CN111649817B, a distributed optical fiber vibration sensor system and a mode identification method thereof are disclosed, wherein after a distributed optical fiber vibration sensor is detected by adopting a coherent detection method to obtain vibration signals at vibration points, intrusion events are classified by an SVM classifier, so that intelligent mode identification of the distributed optical fiber sensor is realized. However, conventional machine learning algorithms require some features to be manually defined in advance, requiring expertise and time consuming. After the deep learning appears, relevant specialists at home and abroad combine the deep learning with the distributed optical fiber sensor, perform pattern recognition on the occurrence of various events, and obtain good recognition effect. In the Chinese patent application publication CN115687994A, an optical fiber event identification and classification method based on a Convolutional Neural Network (CNN) is disclosed, and the method adopts an improved CNN model as a classifier to automatically extract data characteristics, so that the problem of event identification in optical fiber perimeter security is solved at one time and end to end. However, the traditional CNN has the defects of incapability of extracting global features, less learned information and no memory function, so that the model is easy to fit training data, and the actual situation is ignored in the event recognition process, thereby influencing the generalization capability of the model.
Before deep learning develops, a machine learning method based on feature extraction is widely used for identifying various intrusion signals, and although the method can adopt an efficient feature extraction method to improve the identification rate, manually extracted features have strong pertinence, and once the external environment changes, the manually extracted features may not be effective any more. The neural network structure through deep learning can automatically learn more characteristic information, and the strong learning capacity of the neural network structure can enable the neural network structure to express complex functions through a plurality of neural network layers, so that a new calculation space is expanded, and a series of complex problems are solved. Thus, with deep learning of CNN, intrusion signal classification can be effectively performed, and many studies have been focused on event recognition in optical fiber distributed sensors using the technology of neural networks. However, most neural networks use two-dimensional (2-D) convolution kernels to construct the convolution layers, and therefore, in order to process one-dimensional (1-D) sensing signals, the learner always converts the original signals into two-dimensional images or data matrices through time-frequency analysis, but this results in higher computational costs. In order to improve algorithm efficiency, researchers adopt a real-time distributed deep learning network model based on a one-dimensional convolutional neural network (1-D CNN), can capture the characteristics of more one-dimensional time sequences, and has better real-time processing capacity and higher calculation efficiency than the traditional two-dimensional convolutional neural network. However, shallow 1-D CNNs still cannot effectively capture all the information of the 1-D time sequence, and therefore further optimization is required to improve the accuracy of signal identification and reduce the False Alarm Rate (FAR).
Disclosure of Invention
Aiming at the defects of the prior art, the invention provides a perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM.
The invention is realized by the following technical scheme:
the perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM comprises the following steps:
s1, selecting an optical fiber distributed vibration or acoustic wave sensing system based on a phase-sensitive optical time domain reflection technology to build a signal identification system;
s2, adopting a single-mode fiber as a detection optical cable, laying along with the periphery, and setting up at least six events including strong wind, electric drill, climbing, knocking, shaking and sawing and grinding, and collecting data of each event by personnel;
s3, marking the data after the data noise reduction treatment, wherein various events respectively correspond to corresponding labels, dividing a data matrix into single-channel time domain data of each group of independent events, and according to 7:3, dividing the training set and the verification set in proportion to construct a data set;
s4, inputting single-channel time domain data into a CDIL layer, transmitting an output result of the input layer to a first one-dimensional convolution layer, wherein the first one-dimensional convolution layer uses a one-dimensional convolution kernel with the size of 3, the step length is 1, cyclic filling is adopted as a filling model, and the expansion size is 1, so that a convolution result is obtained;
transmitting the result output by the first one-dimensional convolution layer to a second one-dimensional convolution layer, wherein the second one-dimensional convolution layer uses a one-dimensional convolution kernel with the size of 1, and the step length is 1, so as to obtain a convolution result;
transmitting the output result of the second one-dimensional convolution layer to Batch Normalization layer, distributing the output of the upper layer to 0 in the average value and 1 in the variance by Batch Normalization layer, performing normalization operation, and mapping the data x into data xBy mean +.>Sum of variances->Grouping each BatchThe formula is as follows:
activating the output result of the Batch Normalization layer by using a Leaky ReLu function to obtain an output result C (x) of the CDIL layer, and taking the output result C (x) as input of the Bi-LSTM layer;
the Bi-LSTM layer further extracts input characteristics, the input characteristics are input to 2 LSTM networks in a positive sequence and reverse sequence mode, and vectors formed by splicing the 2 output vectors are used as final characteristic expression, so that a Bi-LSTM layer output result Bi (x) is obtained;
fusing the result Bi (x) output by the Bi-LSTM layer with the result C (x) output by the CDIL layer, wherein the calculated output H (x) is as follows:
H(x)＝C(x)+Bi(x)
and transmitting the output results H (x) to a linear layer, giving out preliminary classification logic for each result, then aggregating each logic through an average pooling layer, and finally identifying and classifying the event to be detected through a support vector machine.
Further, in step S2, one end of the detection optical cable is connected with the optical fiber distributed vibration or acoustic wave sensing host, the sampling frequency of data acquisition is 100-10000Hz, the sampling time of each type of event is 5-10min, and the frequency of the event is 3S once.
Further, in step S3, the noise reduction process is performed in a wavelet denoising manner, the signal is decomposed into various scales by wavelet transformation, the wavelet coefficients belonging to the noise are removed at each scale by a method of setting a threshold value, the wavelet coefficients belonging to the signal are retained, and finally the signal data is restored by wavelet inverse transformation.
Further, in step S4, the CDIL layer employs symmetric convolution to exponentially increase the dilation size using dilation convolution with increasing network depth, i.e.:
d l ＝2 l-1 ，
wherein: d, d l For the expansion size of the first convolution layer, a one-dimensional input sequence [ a ] of the first layer convolution is set 1 ，a 2 ，...，a N ]The convolution output at time step t (1. Ltoreq.t. Ltoreq.N) is:
wherein: k is the size of the convolution kernel, and w (l) is the convolution coefficient of the first layer. The fusion of the depth network and the exponential expansion ensures that the receptive field can be effectively and rapidly expanded, the network can be automatically adapted to time sequences of any length, global features can be extracted at one time, and the method has very accurate identification effect on a large-scale data set. In the signal identification process, the model can better capture the integral structure of the signal data, so that the identification accuracy is improved. Needs to be as followsThe CDIL layers are used to achieve a complete receptive field of sequence length. The CDIL layer also buffers boundary effects by cyclic stuffing, handles data movement more robustly, and reduces sensitivity to absolute position information.
In the present invention, the CDIL layer comprises two one-dimensional convolution layers, a Batch Normalization layer, a nonlinear activation layer and an expansion operation module, wherein: the first convolution layer uses a one-dimensional convolution kernel with the size of 3, the step length is 1, and the filling model is circularly filled; the second one-dimensional convolution layer uses a one-dimensional convolution kernel with the size of 1, and the step length is 1; the nonlinear activation layer uses the leak ReLu function as the activation function.
In order to alleviate the internal covariate offset phenomenon and improve the feature extraction capability, batch Normalization (BN) is added, and the BN layer is a regularization technology commonly used in deep neural networks, and can effectively inhibit gradient extinction and gradient explosion, so that the model performance is improved, and the convergence speed is increased. The basic principle of the technology is that the mean and variance of input data are standardized before each layer, then calculation is carried out, and finally the output data are scaled and offset, so that the data distribution is more stable. The ReLU activation function is replaced by the Leaky ReLU, when the input value is smaller than 0, the Leaky ReLU function can calculate a gradient value, the gradient vanishing problem is relieved, small values are captured better, and therefore the performance of the neural network is improved.
Preferably, a spatial attention mechanism is added to the neural network model. And a spatial attention mechanism is added into the network model, so that the representation capability and generalization capability of the model are improved and the performance of the deep neural network is improved by automatically learning the importance of spatial features.
The invention has the beneficial effects that:
1. according to the invention, the cyclic expansion convolutional neural network is introduced into the optical fiber distributed sensor signal recognition, and is different from the deep learning algorithm in the prior art, aiming at the dynamic signal of the optical fiber distributed vibration or sound wave with high sampling rate, the cyclic expansion convolutional neural network can be adaptively expanded to any long time sequence, global characteristics can be extracted at one time, a very accurate recognition effect is achieved on a large-scale data set, and the model can better capture the integral structure of signal data in the optical fiber distributed vibration sensing signal recognition process, so that the recognition accuracy is improved. And adding Batch Normalization layers to accelerate model training and convergence speed, and activating by using a Leaky ReLu function to solve the death ReLU problem of the ReLU neuron during training. In order to better optimize classification, the softmax layer is replaced by a multi-classification SVM classifier, the SVM can effectively process high-dimensional characteristics on a large-scale data set, the generalization capability of a model is improved, the influence of overfitting is weakened, the decision boundary is more stable, and the model can be better fitted in the changed data distribution.
2. The invention carries out feature fusion on the cyclic expansion convolutional neural network and the bidirectional long and short time memory network (Bi-LSTM), wherein the output of the bidirectional long and short time memory network is obtained by simultaneously calculating the sequence information before and after, so the method has more robustness, solves the problem that a single cyclic expansion convolutional neural network model cannot encode the information from back to front to lead to the failure to learn the association between the current feature item and the back feature item, can find the potential relation between data in the optical fiber distributed vibration sensing signal identification process, accurately identifies the feature with strong relativity, improves the identification accuracy, and finally fuses the feature extracted by the cyclic expansion convolutional neural network, prevents feature degradation, and can solve the problem that the model has over fitting and under fitting conditions in the optical fiber distributed vibration sensing signal identification process, thereby causing poor model performance.
3. The CDIL-CNN using time sequence classification task is introduced into optical fiber distributed signal recognition, and a signal recognition method based on feature fusion is provided, the improved CDIL-CNN and Bi-LSTM are fused, and for optical fiber distributed acoustic wave sensing dynamic signals with high sampling rate, unlike the traditional 1-D CNN, the CDIL-CNN can be adaptively expanded to an ultra-long time sequence (with the length of more than or equal to 1000), global features can be extracted at one time, and a model can better capture the integral structure of signal data in the process of optical fiber distributed acoustic wave sensing signal recognition, so that high-precision event classification of a DAS system is realized.
4. The preprocessing method is simple in preprocessing operation, only needs to perform wavelet denoising and data segmentation on the data, does not need to perform a large number of preprocessing operations, ensures the effectiveness and the authenticity of the data, and ensures the accuracy of the identification effect. Training test and comparison are carried out on the network model through an ablation experiment, and the average detection precision of the proposed method on six real disturbance events of strong wind, electric drill, saw grinding, climbing, knocking and shaking under the perimeter security scene can reach more than 99%.
Drawings
FIG. 1 is a flow chart of a perimeter security intrusion signal identification method in the present invention;
FIG. 2 is a flow chart of a signal recognition algorithm according to the present invention;
FIG. 3 is a schematic diagram of a system based on fiber optic distributed acoustic wave sensing;
FIG. 4 is a fence fiber optic cable laying view of the present invention;
FIG. 5 is a feature fusion process of the present invention;
FIG. 6 is a time domain diagram corresponding to a high wind event collected at a perimeter security fence in the present invention;
FIG. 7 is a time domain plot of the invention corresponding to drill events collected at the perimeter security fence;
FIG. 8 is a time domain plot corresponding to a saw and grind event collected at a perimeter security fence in accordance with the present invention;
FIG. 9 is a time domain plot corresponding to a climbing event collected at a perimeter security fence in the present invention;
FIG. 10 is a time domain diagram corresponding to a tapping event collected at a perimeter security fence in the present invention;
FIG. 11 is a time domain plot of the invention corresponding to a wobble event collected at a perimeter security fence.
The figure shows:
the method comprises the steps of collecting a 201 optical fiber distributed acoustic wave sensing signal, denoising 202 wavelets, segmenting and extracting 203 signal data, classifying and manufacturing a data set 204 by labels, inputting a layer 205, circularly expanding a convolution module 206, a 207 bidirectional long-short-time memory network layer 208, a 208 characteristic fusion layer, an 209 average integrated learning module 210, an output layer 211 one-dimensional convolution layer 212Batch Normalization, a 213 expanding operation module 214 nonlinear activation layer 215 linear layer 216 average pooling layer 3 ultra-narrow linewidth laser, 4 acousto-optic modulator, 5 erbium-doped optical fiber amplifier, 6 circulator, 7 photoelectric detector, 8 data acquisition card and 9 upper computer.
Detailed Description
In order to clearly illustrate the technical characteristics of the scheme, the scheme is explained below through a specific embodiment.
A perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM comprises the following steps:
taking a fence as a specific example, the whole flow chart of the perimeter security intrusion signal identification method based on feature fusion is referred to fig. 1. Specific signal processing methods and deep learning network architectures refer to fig. 2.
S1, selecting an optical fiber distributed acoustic wave sensing system based on a phase-sensitive optical time domain reflection technology as a main system for perimeter security intrusion signal identification to build a whole set of system.
The principle of the optical fiber distributed acoustic wave sensing system in this embodiment refers to fig. 3, and the equipment mainly used in the system comprises an ultra-narrow line laser 3, an acousto-optic modulator 4, an erbium-doped optical fiber amplifier 5, a circulator 6, a photoelectric detector 7, a data acquisition card 8 and an upper computer 9. The system of the embodiment adopts a narrow line laser with the line width of 5kHz as a light source, an ultra-narrow line width laser 3 generates a continuous coherent light signal, the continuous coherent light signal is modulated into an optical pulse signal through an acousto-optic modulator 4, an optical pulse sequence with the frequency of 60MHZ is generated, the optical pulse signal is concentrated and amplified by an erbium-doped optical fiber amplifier 5, the amplified optical pulse signal is injected into a detection optical cable through a port I and a port II of an circulator 6, the optical pulse signal generates Rayleigh scattering along the optical cable transmission process, the backward Rayleigh scattering optical signal returns along the optical cable, the backward Rayleigh scattering optical signal enters a photoelectric detector 7 after being received by the port II and the port III of the circulator 6, the signal is acquired by a data acquisition card 8 with the frequency of 100MHz, the introduced phase change information is acquired through phase demodulation, the sound wave and the action signal of vibration on the optical fiber is stored through a personal computer or storage equipment.
S2, adopting a single-mode fiber as a detection optical cable to be laid along with the perimeter fence, and collecting intrusion event samples such as strong wind, electric drill, saw grinding, climbing, knocking, shaking and the like to form event data acquisition. In this embodiment, the single mode fiber is laid in the manner of fig. 4, i.e. one section of fence is laid horizontally and one section of fence is laid vertically, and each section of fence forms three curved loops horizontally and vertically and alternately laid on the fence comprehensively. The transmission optical cable is composed of an optical fiber, a reinforcing layer and a protective layer, after the optical cable is laid, data are collected through the optical fiber distributed acoustic wave sensing system, intrusion events including but not limited to events such as strong wind, electric drills, saw grinding, climbing, knocking, shaking and the like are collected at the fence, the data are stored in a list in a numerical format, and each list stores single collected data of a single event and is stored in a mobile hard disk so as to be convenient to store. The sampling frequency of the optical fiber distributed acoustic wave sensing system is 2500Hz, and each event collecting event is 5-10 min.
S3, performing signal noise reduction processing on the acquired data in a wavelet denoising mode. And extracting time data in the single-mode fiber channel according to parameters of the acquired signals at the perimeter fence, and carrying out noise reduction treatment on the data by a wavelet denoising method. And cutting the denoised data, marking the label, and manufacturing a data set. In this embodiment, the data is divided and extracted by taking 7500 data as a group, each group is about 3s as a complete event, the extracted single-channel time domain data graph refers to fig. 6-11, the signal data is labeled, and the training set and the verification set are divided according to the ratio of 7:3, so as to construct the data set.
S4, inputting single-channel time domain data into the CDIL layer, transmitting an output result of the input layer to a first one-dimensional convolution layer, wherein the first one-dimensional convolution layer uses a one-dimensional convolution kernel with the size of 3, the step length is 1, cyclic filling is adopted as a filling model, and the expansion size is 1, so that a convolution result is obtained.
And transmitting the result output by the first one-dimensional convolution layer to a second one-dimensional convolution layer, wherein the second one-dimensional convolution layer uses a one-dimensional convolution kernel with the size of 1, and the step length is 1, so that a convolution result is obtained.
Transmitting the output result of the second one-dimensional convolution layer to Batch Normalization layer, distributing the output of the upper layer to 0 in the average value and 1 in the variance by Batch Normalization layer, performing normalization operation, and mapping the data x into data xBy mean +.>Sum of variances->Normalizing each Batch, wherein the formula is as follows:
and activating the output result of the Batch Normalization layer by using a leak ReLu function to obtain an activation result.
And the result output by each expansion convolution module is used as the input of a bidirectional long-short-time memory network layer to further extract the characteristics, the bidirectional long-short-time memory network layer is used as an input sequence to be respectively input into 2 LSTM networks in a positive sequence and a reverse sequence to extract the characteristics, and the vectors formed after the 2 output vectors are spliced are used as final characteristic expression.
Fusing a result Bi (x) output by the bidirectional long-short-term memory network layer with a result C (x) output by the cyclic expansion convolutional neural network, wherein the calculated output H (x) is as follows:
H(x)＝C(x)+Bi(x)
the characteristics are fused by a residual network structure, as shown in fig. 5, the output of the CDIL network layer is used as the input of the Bi-LSTM network layer, so as to mine more characteristics, simultaneously retain the characteristics extracted from the CDIL layer, and finally fuse with the Bi-directional characteristics captured by the Bi-LSTM layer, so that the processing can prevent the characteristics from attenuating.
And transmitting the output results H (x) to a linear layer, giving out preliminary classification logic for each result, then aggregating each logic through an average pooling layer, and finally identifying and classifying the event to be detected through a support vector machine.
The algorithm is built for signal identification and classification according to the feature fusion network algorithm structure in fig. 2. Training the manufactured data set and the label in a neural network model based on feature fusion, optimizing parameters by adjusting parameters such as the number of training wheels, the batch size, the learning rate and the like, and storing the weight parameters with the best effect in the model training process for identifying perimeter security intrusion signals.
In the present invention, 7200 single-pass time series data samples of six events were collected, each sample size being 10kx1, and according to 7:3, and the two data sets do not directly intersect.
In order to construct an improved CDIL-BiLSTM network, a PyTorch platform is utilized, which provides rich neural network function interfaces, so that the construction process of the network becomes quicker. In experiments, we used cross entropy as a general loss function for multi-classification problems and employed Adam optimizers to train the network model with a Learning Rate (LR) of 0.001 and a batch size of 64, for a total of 50 iteration periods. The model training process effectively utilizes the powerful computing power of three NVIDIA A100 Tensor Core GPUs, so that the model training efficiency is greatly improved.
The architecture diagram of the improved CDIL-BiLSTM network model is as follows:
the ablation experiment compares the following:
the comparative experiment results are as follows:
of course, the above description is not limited to the above examples, and the technical features of the present invention that are not described may be implemented by or by using the prior art, which is not described herein again; the above examples and drawings are only for illustrating the technical scheme of the present invention and not for limiting the same, and the present invention has been described in detail with reference to the preferred embodiments, and it should be understood by those skilled in the art that changes, modifications, additions or substitutions made by those skilled in the art without departing from the spirit of the present invention and the scope of the appended claims.
Claims (4)
1. A perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM is characterized in that: the method comprises the following steps:
s1, selecting an optical fiber distributed vibration or acoustic wave sensing system based on a phase-sensitive optical time domain reflection technology to build a signal identification system;
s2, adopting a single-mode fiber as a detection optical cable, laying along with the periphery, and setting up at least six events including strong wind, electric drill, climbing, knocking, shaking and sawing and grinding, and collecting data of each event by personnel;
s3, marking the data after the data noise reduction treatment, wherein each event corresponds to a corresponding label, dividing a data matrix into single-channel time domain data of each group of independent events, and dividing a training set and a verification set according to the proportion of 7:3 to construct a data set;
s4, inputting single-channel time domain data into a CDIL layer, transmitting an output result of the input layer to a first one-dimensional convolution layer, wherein the first one-dimensional convolution layer uses a one-dimensional convolution kernel with the size of 3, the step length is 1, cyclic filling is adopted as a filling model, and the expansion size is 1, so that a convolution result is obtained;
transmitting the result output by the first one-dimensional convolution layer to a second one-dimensional convolution layer, wherein the second one-dimensional convolution layer uses a one-dimensional convolution kernel with the size of 1, and the step length is 1, so as to obtain a convolution result;
transmitting the output result of the second one-dimensional convolution layer to Batch Normalization layer, distributing the output of the upper layer to 0 in the average value and 1 in the variance by Batch Normalization layer, performing normalization operation, and mapping the data x into data xBy means of average->Sum of variances->Normalizing each Batch, wherein the formula is as follows:
activating the output result of the Batch Normalization layer by using a Leaky ReLu function to obtain an output result C (x) of the CDIL layer, and taking the output result C (x) as input of the Bi-LSTM layer;
the Bi-LSTM layer further extracts input characteristics, the input characteristics are input to 2 LSTM networks in a positive sequence and reverse sequence mode, and vectors formed by splicing the 2 output vectors are used as final characteristic expression, so that a Bi-LSTM layer output result Bi (x) is obtained;
fusing the result Bi (x) output by the Bi-LSTM layer with the result C (x) output by the CDIL layer, wherein the calculated output H (x) is as follows:
H(x)＝C(x)+Bi(x)
and transmitting the output results H (x) to a linear layer, giving out preliminary classification logic for each result, then aggregating each logic through an average pooling layer, and finally identifying and classifying the event to be detected through a support vector machine.
2. The improved CDIL-Bi-LSTM based perimeter security intrusion signal identification method of claim 1, wherein: in the step S2, one end of the detection optical cable is connected with an optical fiber distributed vibration or sound wave sensing host, the sampling frequency of data acquisition is 100-10000Hz, the sampling time of each type of event is 5-10min, and the frequency of the event is 3S once.
3. The improved CDIL-Bi-LSTM based perimeter security intrusion signal identification method of claim 1, wherein: in step S3, the noise reduction processing is performed in a wavelet denoising manner, signals are decomposed into scales by wavelet transformation, wavelet coefficients belonging to noise are removed at each scale by a method of setting a threshold value, the wavelet coefficients belonging to signals are reserved, and finally, signal data are restored through wavelet inverse transformation.
4. The improved CDIL-Bi-LSTM based perimeter security intrusion signal identification method of claim 1, wherein: in step S4, the CDIL layer employs symmetric convolution, using dilation convolution that increases with network depth, increasing dilation size exponentially, i.e.:
d l ＝2 l-1 ，
wherein: d, d l For the expansion size of the first convolution layer, a one-dimensional input sequence [ a ] of the first layer convolution is set 1 ，a 2 ，...，a N ]The convolution output at time step t (1. Ltoreq.t. Ltoreq.N) is:
wherein: k is the size of the convolution kernel, and w (l) is the convolution coefficient of the first layer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310512549.8A CN116504005B (en) | 2023-05-09 | 2023-05-09 | Perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310512549.8A CN116504005B (en) | 2023-05-09 | 2023-05-09 | Perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116504005A true CN116504005A (en) | 2023-07-28 |
| CN116504005B CN116504005B (en) | 2024-02-20 |
Family
ID=87319992
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310512549.8A Active CN116504005B (en) | 2023-05-09 | 2023-05-09 | Perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116504005B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118051830A (en) * | 2024-04-16 | 2024-05-17 | 齐鲁工业大学(山东省科学院) | Perimeter security intrusion event identification method |
Citations (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110309797A (en) * | 2019-07-05 | 2019-10-08 | 齐鲁工业大学 | Merge the Mental imagery recognition methods and system of CNN-BiLSTM model and probability cooperation |
| KR102021441B1 (en) * | 2019-05-17 | 2019-11-04 | 정태웅 | Method and monitoring camera for detecting intrusion in real time based image using artificial intelligence |
| CN110570613A (en) * | 2019-09-18 | 2019-12-13 | 广州亓行智能科技有限公司 | Fence vibration intrusion positioning and mode identification method based on distributed optical fiber system |
| CN111104891A (en) * | 2019-12-13 | 2020-05-05 | 天津大学 | Composite characteristic optical fiber sensing disturbing signal mode identification method based on BiLSTM |
| US20200320438A1 (en) * | 2019-04-05 | 2020-10-08 | Zscaler, Inc. | Prudent ensemble models in machine learning with high precision for use in network security |
| CN111970309A (en) * | 2020-10-20 | 2020-11-20 | 南京理工大学 | Spark Internet of vehicles based combined deep learning intrusion detection method and system |
| CN113132399A (en) * | 2021-04-23 | 2021-07-16 | 中国石油大学(华东) | Industrial control system intrusion detection method based on time convolution network and transfer learning |
| CN114199362A (en) * | 2021-12-17 | 2022-03-18 | 齐鲁工业大学 | Distributed optical fiber vibration sensor mode identification method |
| CN114510960A (en) * | 2021-12-28 | 2022-05-17 | 齐鲁工业大学 | Method for recognizing distributed optical fiber sensor system mode |
| CN114510992A (en) * | 2021-12-27 | 2022-05-17 | 上海梦象智能科技有限公司 | Equipment switch state detection method based on deep learning |
| CN114841202A (en) * | 2022-04-24 | 2022-08-02 | 浙江浙能天然气运行有限公司 | Deep learning mode identification method for distributed optical fiber pipeline intrusion detection |
| WO2023010658A1 (en) * | 2021-08-06 | 2023-02-09 | 大连理工大学 | Time dilated convolutional network-based method for alerting of rotating stall in compressor |
| CN115759395A (en) * | 2022-11-15 | 2023-03-07 | 广东电网有限责任公司 | Training of photovoltaic detection model, detection method of photovoltaic power generation and related device |
| CN115859190A (en) * | 2022-11-12 | 2023-03-28 | 上海梦象智能科技有限公司 | Non-invasive household electrical classification method based on causal relationship |
| US11630996B1 (en) * | 2017-06-23 | 2023-04-18 | Virginia Tech Intellectual Properties, Inc. | Spectral detection and localization of radio events with learned convolutional neural features |
-
2023
- 2023-05-09 CN CN202310512549.8A patent/CN116504005B/en active Active
Patent Citations (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11630996B1 (en) * | 2017-06-23 | 2023-04-18 | Virginia Tech Intellectual Properties, Inc. | Spectral detection and localization of radio events with learned convolutional neural features |
| US20200320438A1 (en) * | 2019-04-05 | 2020-10-08 | Zscaler, Inc. | Prudent ensemble models in machine learning with high precision for use in network security |
| KR102021441B1 (en) * | 2019-05-17 | 2019-11-04 | 정태웅 | Method and monitoring camera for detecting intrusion in real time based image using artificial intelligence |
| CN110309797A (en) * | 2019-07-05 | 2019-10-08 | 齐鲁工业大学 | Merge the Mental imagery recognition methods and system of CNN-BiLSTM model and probability cooperation |
| CN110570613A (en) * | 2019-09-18 | 2019-12-13 | 广州亓行智能科技有限公司 | Fence vibration intrusion positioning and mode identification method based on distributed optical fiber system |
| CN111104891A (en) * | 2019-12-13 | 2020-05-05 | 天津大学 | Composite characteristic optical fiber sensing disturbing signal mode identification method based on BiLSTM |
| US20220217170A1 (en) * | 2020-10-20 | 2022-07-07 | Nanjing University Of Science And Technology | Intrusion detection method and system for internet of vehicles based on spark and deep learning |
| CN111970309A (en) * | 2020-10-20 | 2020-11-20 | 南京理工大学 | Spark Internet of vehicles based combined deep learning intrusion detection method and system |
| CN113132399A (en) * | 2021-04-23 | 2021-07-16 | 中国石油大学(华东) | Industrial control system intrusion detection method based on time convolution network and transfer learning |
| WO2023010658A1 (en) * | 2021-08-06 | 2023-02-09 | 大连理工大学 | Time dilated convolutional network-based method for alerting of rotating stall in compressor |
| CN114199362A (en) * | 2021-12-17 | 2022-03-18 | 齐鲁工业大学 | Distributed optical fiber vibration sensor mode identification method |
| CN114510992A (en) * | 2021-12-27 | 2022-05-17 | 上海梦象智能科技有限公司 | Equipment switch state detection method based on deep learning |
| CN114510960A (en) * | 2021-12-28 | 2022-05-17 | 齐鲁工业大学 | Method for recognizing distributed optical fiber sensor system mode |
| CN114841202A (en) * | 2022-04-24 | 2022-08-02 | 浙江浙能天然气运行有限公司 | Deep learning mode identification method for distributed optical fiber pipeline intrusion detection |
| CN115859190A (en) * | 2022-11-12 | 2023-03-28 | 上海梦象智能科技有限公司 | Non-invasive household electrical classification method based on causal relationship |
| CN115759395A (en) * | 2022-11-15 | 2023-03-07 | 广东电网有限责任公司 | Training of photovoltaic detection model, detection method of photovoltaic power generation and related device |
Non-Patent Citations (6)
| Title |
|---|
| JINGCHENG HUANG: "A Fiber Vibration Signal Recognition Method Based on CNN-CBAM-LSTM", APPLIED SCIENCES, vol. 17, no. 12, pages 8478 - 8489 * |
| LEI CHENG: "Classification of long sequential data using circular dilated convolutional neural networks", NEUROCOMPUTING, vol. 2023, no. 518, pages 50 - 59 * |
| 周子纯: "基于卷积长短期记忆全连接深度神经网络的光纤振动传感事件识别", 光学学报, vol. 41, no. 13, pages 139 - 147 * |
| 张学兵: "基于CNN-LSTM 混合神经网络的高速铁路地震响应预测", 湘潭大学学报, pages 1 - 13 * |
| 王昌: "分布式光纤声波地震波勘探技术", 山东科学, vol. 34, no. 4, pages 1 - 8 * |
| 黄莺莺: "基于深度学习的声纹识别方法研究", 中国优秀硕士学位论文全文数据库信息科技辑, no. 8, pages 110 - 136 * |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118051830A (en) * | 2024-04-16 | 2024-05-17 | 齐鲁工业大学(山东省科学院) | Perimeter security intrusion event identification method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116504005B (en) | 2024-02-20 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108932480B (en) | Distributed optical fiber sensing signal feature learning and classifying method based on 1D-CNN | |
| CN109214349B (en) | Object detection method based on semantic segmentation enhancement | |
| CN110995339A (en) | Method for extracting and identifying time-space information of distributed optical fiber sensing signal | |
| CN101893704A (en) | Rough set-based radar radiation source signal identification method | |
| CN116504005B (en) | Perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM | |
| CN111859010B (en) | Semi-supervised audio event identification method based on depth mutual information maximization | |
| CN116579616B (en) | Risk identification method based on deep learning | |
| CN112508014A (en) | Improved YOLOv3 target detection method based on attention mechanism | |
| CN110599459A (en) | Underground pipe network risk assessment cloud system based on deep learning | |
| CN112766223B (en) | Hyperspectral image target detection method based on sample mining and background reconstruction | |
| CN113920472A (en) | Unsupervised target re-identification method and system based on attention mechanism | |
| CN112149612A (en) | Marine organism recognition system and recognition method based on deep neural network | |
| Mathias et al. | Deep Neural Network Driven Automated Underwater Object Detection. | |
| CN115859186A (en) | Distributed optical fiber sensing event identification method and system based on Grarami angular field | |
| CN114187546B (en) | Combined action recognition method and system | |
| Yaman et al. | Image processing and machine learning‐based classification method for hyperspectral images | |
| CN113034543B (en) | 3D-ReID multi-target tracking method based on local attention mechanism | |
| Wang et al. | FPA-DNN: a forward propagation acceleration based deep neural network for ship detection | |
| CN115147727A (en) | Method and system for extracting impervious surface of remote sensing image | |
| Wang et al. | Sonar Objective Detection Based on Dilated Separable Densely Connected CNNs and Quantum‐Behaved PSO Algorithm | |
| Bhandari et al. | An innovative remote sensing image retrieval techniques based on haar wavelet-LTRP and ANFIS | |
| CN114140524A (en) | Closed loop detection system and method for multi-scale feature fusion | |
| Yang et al. | Distributed Optical Fiber Sensing Event Recognition Based on Markov Transition Field and Knowledge Distillation | |
| He et al. | Snowflakes removal for single image based on model pruning and generative adversarial network | |
| CN116863251B (en) | Distributed optical fiber sensing disturbance recognition method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |